Add xe2 block_fp8 grouped_gemm kernel

[frost-intel]

Purpose

Add block-wise FP8 (W8A16) weight quantization support to the Xe2 grouped GEMM that backs the fused-MoE op. Weights are stored in FP8 (e4m3/e5m2) with a 128×128 block scale grid and dequantized on the fly against BF16/FP16 activations, complementing the existing per-tensor FP8 path. This matches the block-scaled FP8 checkpoints in DeepSeek-V2/V3, Qwen3-MoE, etc.

Test Plan

• Correctness — block-FP8 output vs a dequantized (fp8 → x_dtype × scale) reference MoE:

  pytest tests/fused_moe/test_grouped_gemm.py

  Also exercised via the benchmark's calculate_diff against ref_fused_moe across model shapes, both activation dtypes (bf16/fp16), both weight dtypes (e4m3/e5m2), and with/without bias.
• Performance — per-stage (remap / gemm1 / gemm2 / gather) latency + TFLOPS sweep over representative MoE shapes:

  python benchmark/benchmark_cutlass_fused_moe.py

Test Result

Tests were executed on PVC Max 1550.

• Correctness: all configs pass torch.testing.assert_close within the per-dtype rtol/atol (block-FP8, per-tensor FP8, and BF16 baselines).
• Performance: block-FP8 is ~10% slower than per-tensor FP8, with the worst gap at large M, and is within a few percent of per-tensor at small/decode M.

DeepSeek-V3

My objective is to improve DSv3 performance. For those kernels specifically, we achieve a mean of 1.04x

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[Copilot]

Pull request overview

This PR adds block-wise FP8 (W8A16) support for the Xe2 grouped GEMM used by the fused MoE op, enabling 128×128 (K×N) block scale grids to match block-scaled FP8 checkpoints.

Changes:

• Extends the CUTLASS grouped GEMM interface end-to-end with an is_B_fp8block flag and block-scale validation.
• Implements a new Xe2 device kernel path (xe_gemm_blockfp8) that applies per-128×128 block FP32 scales for FP8 weights.
• Adds correctness/perf benchmarking plumbing and a new unit test covering block-FP8 grouped GEMM.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
vllm_xpu_kernels/fused_moe_interface.py	Plumbs is_B_fp8block through Python fused-MoE calls into the XPU op.
tests/fused_moe/test_grouped_gemm.py	Adds a new block-FP8 grouped GEMM correctness test.
csrc/xpu/torch_bindings.cpp	Updates the Torch library schema to include is_B_fp8block.
csrc/xpu/ops.h	Updates the C++ op declaration with the new boolean parameter.
csrc/xpu/grouped_gemm/xe_2/grouped_gemm_xe2.hpp	Adds compile-time IsBlockScale handling and dispatch to the block-FP8 GEMM path.
csrc/xpu/grouped_gemm/xe_2/grouped_gemm_xe2.h	Extends Xe2 grouped GEMM API to accept is_B_fp8block.
csrc/xpu/grouped_gemm/xe_2/grouped_gemm_xe2.cpp	Forwards is_B_fp8block into the Xe2 implementation.
csrc/xpu/grouped_gemm/xe_2/grouped_gemm_xe2_interface.hpp	Adds block-FP8 argument validation and launch plumbing (IsBlockScale=true).
csrc/xpu/grouped_gemm/xe_2/gemm_xe2.hpp	Implements xe_gemm_blockfp8 device kernel for block-scaled FP8 weights.
csrc/xpu/grouped_gemm/grouped_gemm_interface.h	Extends the top-level grouped GEMM interface signature with is_B_fp8block.
csrc/xpu/grouped_gemm/grouped_gemm_interface.cpp	Wires is_B_fp8block into the Xe2 path and rejects it for XE default kernel.
benchmark/src/get_model_config.py	Adds block-FP8 variations to correctness/perf config generation with validity filtering.
benchmark/src/fused_moe_interface_.py	Adds is_block_fp8 flag and forwards it to grouped GEMM calls.
benchmark/benchmark_cutlass_fused_moe.py	Adds block-FP8 data generation, diff checking, and benchmark parameterization.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[xinyu-intel]

thanks for the contribution. I tend to merge this first. #376

[jikunshang]

please rebase and fix conflicts.